Cord
cord is a coordination protocol for trees of Claude Code agents. It enables a parent agent to spawn child agents, delegate sub-tasks to each, and collect structured results — bringing a supervisor pattern to multi-agent Claude Code pipelines. State sharing and task handoff between agents follow a defined protocol rather than ad hoc prompt chaining.
Claude Code is a single-agent tool by default: one model, one conversation, one tool set. Teams building more complex agentic workflows today coordinate multiple Claude Code instances through informal conventions — shared files, manual prompt hand-offs, or bespoke orchestration scripts. cord proposes a standard coordination layer: a parent agent decomposes a large task, spawns specialised child agents for each piece, and receives results back through a consistent protocol. This mirrors the supervisor pattern common in distributed systems but applied to LLM agent trees.
Placed in Assess at inner position because multi-agent Claude Code orchestration is an unsolved and actively explored problem. cord is early and experimental, but it addresses a real gap: the absence of a standard protocol for structured agent coordination within Claude Code. The inner position reflects that developers building multi-agent Claude Code pipelines should be evaluating cord now, alongside other emerging approaches such as the subagent patterns in the Claude Agent SDK. Teams doing single-agent Claude Code work have no immediate need to assess it.
